Replace with LTX-2.5 Docker Space app (text_to_video / i2v + narration audio)

.dockerignore

@@ -0,0 +1,6 @@
+.git
+.gitignore
+.gitattributes
+__pycache__
+*.pyc
+.DS_Store

.gitattributes

@@ -32,4 +32,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.xz filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
-*tfevents* filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -1,11 +1,6 @@
-# Python-generated files
-__pycache__/
-*.py[oc]
-build/
-dist/
-wheels/
-*.egg-info
-
-# Virtual environments
-.venv
-.env
+# Local model / cache artifacts
+*.gguf
+*.safetensors
+/tmp/
+.cache/
+__pycache__/

.python-version

@@ -1 +0,0 @@
-3.12

Dockerfile

@@ -1,21 +1,22 @@
+FROM python:3.12-slim
 
-FROM python:3.10
+ENV PYTHONDONTWRITEBYTECODE=1 \
+    PYTHONUNBUFFERED=1 \
+    PIP_NO_CACHE_DIR=1
 
-# Pehle root user hai by default, to yahan packages install karo
-RUN apt-get update && apt-get install -y libgl1 libglib2.0-0
+WORKDIR /app
 
-# Ab user add karo aur switch karo
-RUN useradd -m -u 1000 user
-USER user
-ENV PATH="/home/user/.local/bin:$PATH"
+RUN apt-get update \
+    && apt-get install -y --no-install-recommends ffmpeg git \
+    && rm -rf /var/lib/apt/lists/*
 
-WORKDIR /app
+COPY requirements.txt .
+RUN pip install --no-cache-dir torch --index-url https://download.pytorch.org/whl/cu124 \
+    && pip install --no-cache-dir -r requirements.txt \
+    && pip install --no-cache-dir gradio spaces huggingface_hub
 
+COPY . .
 
-COPY --chown=user ./requirements.txt requirements.txt
-RUN pip install --no-cache-dir --upgrade -r requirements.txt
-RUN python -m spacy download en_core_web_lg
-RUN python -c "from doctr.models import ocr_predictor; ocr_predictor(pretrained=True)"
+EXPOSE 7860
 
-COPY --chown=user . /app
-CMD ["uvicorn", "app:app", "--host", "0.0.0.0", "--port", "7860"]
+CMD ["python", "app.py"]

README.md

@@ -1,9 +1,14 @@
 ---
-title: DoctorXpert Backend
-emoji: 🩺
-colorFrom: yellow
-colorTo: red
+title: Video Creator
+emoji: 🔥
+colorFrom: green
+colorTo: blue
 sdk: docker
 app_file: app.py
 pinned: false
 ---
+
+Docker-hosted LTX-2.5 Space for the AI Shorts Factory backend. Four Gradio
+endpoints via `/call/{fn}` (text_to_video, text_to_video_av, image_to_video,
+image_to_video_av). Requires a GPU tier — the free CPU tier (2 vCPU/16GB)
+boots but generation OOMs.

app.py

@@ -1,57 +1,355 @@
-from fastapi import FastAPI, UploadFile , File , Request
-from fastapi.middleware.cors import CORSMiddleware
-from agents import Runner
-from logging import getLogger
-from reportanalysis import Report_Agent, extract_text
-from chatbot import get_health_response 
-app = FastAPI()
-log = getLogger()
-
-app.add_middleware(
-    CORSMiddleware,
-    allow_origins=["*"],
-    allow_credentials=True,
-    allow_methods=["*"],
-    allow_headers=["*"],
+"""LTX-2.5 Space — ZeroGPU inference API for the AI Shorts Factory backend.
+
+Runs the distilled LTX-2.5 GGUF (Q4_K_M) on zero GPU as a Gradio demo. Four
+endpoints are exposed through Gradio's `/call/{fn}` protocol and consumed by
+the backend provider `app.ai.visuals.ltx`:
+
+    text_to_video(prompt, width, height, num_frames, enhance_prompt)
+    text_to_video_av(...)            # video + synchronized 48kHz narration audio
+    image_to_video(prompt, image_url, width, height, num_frames, enhance_prompt)
+    image_to_video_av(...)           # video + synchronized narration audio
+
+Every function returns ``(video_path, first_frame_path)``. Scene 2..N of the
+pipeline call image_to_video with the last frame URL of the previous clip so
+the subject stays continuous (last-frame chaining).
+
+The backend always sends an already-composed prompt (quoted dialogue for the
+exact script words) and passes ``enhance_prompt=False``; the prompt enhancer
+is a hard no-op here so no rewrite happens.
+"""
+
+from __future__ import annotations
+
+import os
+import subprocess
+import threading
+
+# ZeroGPU rule #1: `import spaces` must precede any CUDA-touching import
+# (torch etc.) — it monkey-patches torch.cuda at import time.
+import spaces  # noqa: E402
+
+import numpy as np
+import torch
+
+HF_TOKEN = os.environ.get("HF_TOKEN", "")
+
+# Direct download URL of the distilled GGUF transformer (~15GB). The backend
+# only needs LTX_SPACE_URL; this key is configured in the Space.
+LTX_MODEL_URL = os.environ.get(
+    "LTX_MODEL_URL",
+    "https://huggingface.co/realrebelai/LTX-2.5_GGUFs/resolve/main/LTX-2.5-Distilled-Q4_K_M.gguf",
 )
+PIPELINE_ID = "Lightricks/LTX-2.5-Diffusers"
+SEED = int(os.environ.get("LTX_SEED", "0"))
+FPS = int(os.environ.get("LTX_FPS", "24"))
+
+# Distilled model: guidance is baked into the weights, few-step schedule.
+INFERENCE_STEPS = int(os.environ.get("LTX_STEPS", "8"))
+GUIDANCE_SCALE = float(os.environ.get("LTX_GUIDANCE", "1.0"))
+NEGATIVE_PROMPT = "low quality, blurry, distorted, watermark"
+# Honest worst-case GPU wall-time per generation. ZeroGPU validates this
+# after a 1.5x multiplier and sits a ~60-120s continuous execution cap per
+# call on the free tier — 120 keeps us inside both (120 * 1.5 = 180 < 300
+# cap, and matches the real per-call window after the model is resident).
+GPU_DURATION = int(os.environ.get("LTX_GPU_DURATION", "120"))
+
+_lock = threading.Lock()
+_model = None
+
+
+# ----------------------------------------------------------------- model load
+
+
+def _download_gguf(dest: str) -> str:
+    import httpx
+
+    if LTX_MODEL_URL.startswith("http"):
+        if os.path.exists(dest):
+            return dest
+        with httpx.stream("GET", LTX_MODEL_URL, follow_redirects=True, timeout=600) as r:
+            r.raise_for_status()
+            with open(dest, "wb") as fh:
+                for chunk in r.iter_bytes(chunk_size=1 << 20):
+                    fh.write(chunk)
+        return dest
+    from huggingface_hub import hf_hub_download
+
+    # Support "repo_id:filename" shorthand.
+    repo, _, filename = LTX_MODEL_URL.partition(":")
+    return hf_hub_download(repo, filename or "LTX-2.5-Distilled-Q4_K_M.gguf", token=HF_TOKEN)
+
 
-@app.get("/")
-async def root():
-    return {"message": "Welcome to the Medical Report Analysis API"}
+def _build_transformer(gguf_path: str):
+    from diffusers import AutoModel
+    from diffusers.utils import GGUFQuantizationConfig
+
+    kwargs = dict(
+        quantization_config=GGUFQuantizationConfig(compute_dtype=torch.bfloat16),
+        dtype=torch.bfloat16,
+    )
+    try:
+        return AutoModel.from_single_file(gguf_path, config=PIPELINE_ID, **kwargs)
+    except Exception as exc:  # config may live elsewhere; trust the GGUF KV metadata
+        print(f"[ltx] from_single_file with config failed ({exc}), retrying without")
+        return AutoModel.from_single_file(gguf_path, **kwargs)
+
+
+def _load_model() -> dict:
+    global _model
+    if _model is not None:
+        return _model
+
+    with _lock:
+        if _model is not None:
+            return _model
+
+        print("[ltx] downloading GGUF transformer…")
+        gguf_path = os.environ.get("LTX_GGUF_CACHE", "/tmp/ltx-model.gguf")
+        gguf_path = _download_gguf(gguf_path)
+
+        print("[ltx] building quantized transformer…")
+        transformer = _build_transformer(gguf_path)
+
+        from diffusers import LTX2ImageToVideoPipeline, LTX2Pipeline
+
+        built = {
+            "transformer": transformer,
+        }
+        try:
+            built["t2v"] = LTX2Pipeline.from_pretrained(
+                PIPELINE_ID, transformer=transformer, torch_dtype=torch.bfloat16
+            )
+            built["i2v"] = LTX2ImageToVideoPipeline.from_pretrained(
+                PIPELINE_ID, transformer=transformer, torch_dtype=torch.bfloat16
+            )
+        except Exception as exc:  # pragma: no cover - component layout differences
+            raise RuntimeError(
+                "[ltx] pipeline build failed - check Lightricks/LTX-2.5-Diffusers "
+                f"component layout. {exc}"
+            ) from exc
+
+        for pipe in (built["t2v"], built["i2v"]):
+            pipe.enable_model_cpu_offload()
+        _model = built
+        print("[ltx] model ready")
+        return _model
 
-@app.post("/chatbot")
-async def health_info(request: Request):
-    data = await request.json()
-    msg = data.get("message")
-    result = await get_health_response(msg)  # <- await here
-    result.replace("\n/", "<br>")
-    return {"response": result}
 
-@app.post("/report-analysis")
-async def upload_file(file: UploadFile = File(...)):
+# ZeroGPU rule #2: load the model at module scope so weights are packed once
+# at startup and resident for every forked worker, instead of per-call.
+# `_generate` still guards with `_load_model()` so early calls block on the
+# lock until this background warm-up finishes.
+_warmup = threading.Thread(target=_load_model, name="ltx-warmup", daemon=True)
+_warmup.start()
+
+
+# ----------------------------------------------------------------- generation
+
+
+def _try_call(pipe, **kwargs):
+    """Call the pipeline, disabling the (Gemma-4 based) prompt enhancer."""
     try:
-        pdf = file.filename.lower().endswith('.pdf')
-        doc = file.filename.lower().endswith('.docx')
-        content = await file.read()
-        text = extract_text(content, pdf, doc)
-        result = await Runner.run(
-            Report_Agent, 
-           f"""Please analyze the uploaded medical report image extrected text :
+        return pipe(enhance_prompt=False, **kwargs)
+    except TypeError:
+        return pipe(**kwargs)
+
+
+def _extract_frames(output) -> list:
+    frames = output.frames
+    if isinstance(frames[0], (list, tuple)):
+        return list(frames[0])
+    return list(frames)
+
+
+def _extract_audio(output):
+    audio = getattr(output, "audio", None)
+    if audio is None:
+        return None
+    if isinstance(audio, (list, tuple)):
+        arr, sr = audio[0], audio[1] if len(audio) > 1 else 48000
+    else:
+        arr, sr = audio, 48000
+    if torch.is_tensor(arr):
+        arr = arr.detach().float().cpu().numpy()
+    arr = np.asarray(arr)
+    if arr.ndim == 3:  # (batch, samples, channels)
+        arr = arr[0]
+    if arr.ndim == 2:  # downmix to mono for a clean narration bed
+        arr = arr.mean(axis=1)
+    return arr.astype(np.float32), int(sr)
+
 
-{text}
+def _mux_audio(video_path, frames, fps, audio) -> str:
+    from diffusers.utils import export_to_video
 
-Do not provide any analysis before calling the tool.
+    export_to_video(frames, video_path, fps=fps)
+    if audio is None:
+        return video_path
 
-Once the text is extracted, continue with step-by-step medical analysis and return the final output strictly in JSON format.
+    import imageio_ffmpeg
+    from scipy.io import wavfile
 
-    """,
-    context=content,
-    
+    arr, sr = audio
+    wav_path = video_path.with_suffix(".wav")
+    wavfile.write(str(wav_path), sr, arr)
+    ffmpeg = imageio_ffmpeg.get_ffmpeg_exe()
+    muxed = str(video_path.with_suffix("_mux.mp4"))
+    subprocess.run(
+        [
+            ffmpeg, "-y",
+            "-i", str(video_path),
+            "-i", str(wav_path),
+            "-map", "0:v", "-map", "1:a",
+            "-c:v", "copy", "-c:a", "aac", "-shortest",
+            muxed,
+        ],
+        check=True,
+        capture_output=True,
+    )
+    return muxed
+
+
+def _load_image(image_url: str):
+    if not image_url:
+        return None
+    import base64
+    import io
+
+    from PIL import Image, ImageOps
+
+    if image_url.startswith("data:"):
+        encoded = image_url.partition(",")[2]
+        return Image.open(io.BytesIO(base64.b64decode(encoded))).convert("RGB")
+
+    import httpx
+
+    data = httpx.get(image_url, timeout=120).content
+    image = Image.open(io.BytesIO(data)).convert("RGB")
+    return ImageOps.exif_transpose(image)
+
+
+def _generate(prompt, image_url, width, height, num_frames, enhance_prompt, with_audio):
+    if enhance_prompt:
+        # Never let the enhancer rewrite the scripted dialogue.
+        enhance_prompt = False
+
+    num_frames = max(9, int(num_frames))
+    num_frames = 1 + (num_frames - 1 - ((num_frames - 1) % 8))  # 1 + 8n
+    width = int(width) // 32 * 32
+    height = int(height) // 32 * 32
+
+    model = _load_model()
+    generator = torch.Generator(device="cuda").manual_seed(SEED)
+
+    image = _load_image(image_url)
+    common = dict(
+        prompt=prompt,
+        negative_prompt=NEGATIVE_PROMPT,
+        width=width,
+        height=height,
+        num_frames=num_frames,
+        num_inference_steps=INFERENCE_STEPS,
+        guidance_scale=GUIDANCE_SCALE,
+        generator=generator,
+    )
+    if image is not None:
+        output = _try_call(model["i2v"], image=image, **common)
+    else:
+        output = _try_call(model["t2v"], **common)
+
+    frames = _extract_frames(output)
+    audio = _extract_audio(output) if with_audio else None
+
+    video_path = f"/tmp/gradio/ltx_{threading.get_ident()}.mp4"
+    muxed = _mux_audio(video_path, frames, FPS, audio)
+
+    first_frame = f"/tmp/gradio/ltx_{threading.get_ident()}_first.png"
+    frames[0].save(first_frame)
+    return muxed, first_frame
+
+
+# ------------------------------------------------------------------ gradio UI
+
+
+def text_to_video(prompt, width, height, num_frames, enhance_prompt=False):
+    return _generate(prompt, "", width, height, num_frames, enhance_prompt, with_audio=False)
+
+
+def text_to_video_av(prompt, width, height, num_frames, enhance_prompt=False):
+    return _generate(prompt, "", width, height, num_frames, enhance_prompt, with_audio=True)
+
+
+def image_to_video(prompt, image_url, width, height, num_frames, enhance_prompt=False):
+    return _generate(prompt, image_url, width, height, num_frames, enhance_prompt, with_audio=False)
+
+
+def image_to_video_av(prompt, image_url, width, height, num_frames, enhance_prompt=False):
+    return _generate(prompt, image_url, width, height, num_frames, enhance_prompt, with_audio=True)
+
+
+import gradio as gr  # noqa: E402
+
+t2v_fns = [spaces.GPU(duration=GPU_DURATION)(text_to_video), spaces.GPU(duration=GPU_DURATION)(text_to_video_av)]
+i2v_fns = [spaces.GPU(duration=GPU_DURATION)(image_to_video), spaces.GPU(duration=GPU_DURATION)(image_to_video_av)]
+
+with gr.Blocks(title="LTX-2.5 Shorts Space") as demo:
+    gr.Markdown(
+        "# LTX-2.5 Shorts Space\n"
+        "ZeroGPU inference for the AI Shorts Factory backend. "
+        "Portrait 9:16 clips (~4s) with optional synchronized narration audio. "
+        "The prompt enhancer stays **off** so quoted dialogue is spoken exactly."
+    )
+
+    width = gr.Slider(256, 768, value=544, step=32, label="Width (÷32)")
+    height = gr.Slider(256, 1408, value=960, step=32, label="Height (÷32)")
+    num_frames = gr.Slider(9, 801, value=97, step=8, label="Frames (8n+1)")
+    enhance = gr.Checkbox(value=False, label="Enhance prompt (kept off)")
+    video_out = gr.Video(label="Clip")
+    frame_out = gr.Image(label="First frame (for chaining)")
+
+    with gr.Tab("Text → Video"):
+        t2v_prompt = gr.Textbox(label="Prompt", lines=4)
+        t2v_btn = gr.Button("Generate (video only)")
+        t2v_btn.click(
+            t2v_fns[0],
+            inputs=[t2v_prompt, width, height, num_frames, enhance],
+            outputs=[video_out, frame_out],
         )
-        
-        print(result.final_output)
-        return {"result": result.final_output.model_dump()}
-    except Exception as e:
-        return {"error": str(e)}
 
+    with gr.Tab("Text → Video + Audio"):
+        t2va_prompt = gr.Textbox(label="Prompt", lines=4)
+        t2va_btn = gr.Button("Generate (with narration)")
+        t2va_btn.click(
+            t2v_fns[1],
+            inputs=[t2va_prompt, width, height, num_frames, enhance],
+            outputs=[video_out, frame_out],
+        )
+
+    with gr.Tab("Image → Video"):
+        i2v_prompt = gr.Textbox(label="Prompt", lines=4)
+        i2v_image = gr.Textbox(
+            label="First-frame image URL (last frame from the previous clip)",
+            value="",
+        )
+        i2v_btn = gr.Button("Generate (video only)")
+        i2v_btn.click(
+            i2v_fns[0],
+            inputs=[i2v_prompt, i2v_image, width, height, num_frames, enhance],
+            outputs=[video_out, frame_out],
+        )
+
+    with gr.Tab("Image → Video + Audio"):
+        i2va_prompt = gr.Textbox(label="Prompt", lines=4)
+        i2va_image = gr.Textbox(
+            label="First-frame image URL (last frame from the previous clip)",
+            value="",
+        )
+        i2va_btn = gr.Button("Generate (with narration)")
+        i2va_btn.click(
+            i2v_fns[1],
+            inputs=[i2va_prompt, i2va_image, width, height, num_frames, enhance],
+            outputs=[video_out, frame_out],
+        )
 
+if __name__ == "__main__":
+    demo.queue(default_concurrency_limit=1).launch(server_name="0.0.0.0")

chatbot.py

@@ -1,146 +0,0 @@
-# chatbot.py
-from agents import (Agent,
-                    RunConfig,
-                    Runner,
-                    OpenAIChatCompletionsModel,
-                    AsyncOpenAI,
-                    model_settings,
-                    function_tool,
-                    set_tracing_disabled,
-                    enable_verbose_stdout_logging)
-from dotenv import load_dotenv
-import os
-import requests
-set_tracing_disabled(disabled=True)
-enable_verbose_stdout_logging()
-
-load_dotenv()
-api_key = os.getenv("GEM_API_KEY")
-
-external_client = AsyncOpenAI(
-    api_key=api_key,
-    base_url="https://generativelanguage.googleapis.com/v1beta/openai/",
-)
-
-model = OpenAIChatCompletionsModel(
-    model="gemini-2.5-flash",
-    openai_client=external_client
-)
-
-config = RunConfig(
-    model=model,
-    model_provider=external_client,
-    tracing_disabled=True
-)
-
-@function_tool
-async def get_info_about_health(query:str) -> str:
-    """Fetch health information from web based on the query.
-       That helps to provide accurate medical advice."""
-    url = f"https://wsearch.nlm.nih.gov/ws/query?db=healthTopics&term={query}"
-    responce = requests.get(url)
-
-    return responce.text
-
-agent: Agent = Agent(
-    name="Doctor",
-    instructions="""
-You are DrXpert, a professional, confident, and caring AI medical expert.
-Your job is to analyze symptoms, explain possible common causes, and give safe, evidence-based general health guidance — like a real doctor in an initial consultation.
-
-Rules:
-
-Be calm, clear, and empathetic.
-
-Use verified medical knowledge, but never confirm a diagnosis.
-
-Suggest only safe OTC medicines (Paracetamol, Panadol, ORS, Antacid).
-
-❌ Never mention antibiotics, injections, or prescription drugs.
-
-If serious or unclear → “This may need urgent medical attention. Please visit a nearby hospital.”
-
-If unsure → “I’m not completely sure; a doctor’s check-up is best.”
-
-Non-health questions → “I’m designed for health topics only.”
-
-Formatting:
-
-Always reply in clear sections with line breaks.
-
-Use numbered headings (1️⃣, 2️⃣, 3️⃣ …) or bullet points for clarity.
-
-Each section (Causes, Medicine, Precautions, Remedies, Closing) should be on a separate line.
-
-Write in short Urdu-English sentences (Hinglish style).
-
-Response Format:
-1️⃣ Possible Causes: 1–3 short causes
-2️⃣ Safe Medicine: Only mild OTC suggestion
-3️⃣ Precautions: 2–3 points
-4️⃣ Home Remedies: 1–2 simple tips (Urdu + English)
-5️⃣ Kind Closing: Warm, caring line like “Insha’Allah you’ll feel better soon ❤️”
-
-Tone:
-Professional yet warm — like a senior doctor talking gently to a patient. Avoid medical jargon.
-
-✅ Example (Correctly Formatted Reply):
-
-User: “I feel numbness in my leg.”
-DrXpert:
-1️⃣ Possible Causes:
-
-Sitting too long in one position (Aik hi position mein der tak baithna)
-
-Poor blood circulation (Khoon ki gardish mein kami)
-
-Nerve compression (Asab par pressure)
-
-2️⃣ Safe Medicine:
-
-Gently massage the area. Koi pain relief balm laga sakte hain.
-
-3️⃣ Precautions:
-
-Move every 20–30 minutes.
-
-Maintain good posture while sitting.
-
-4️⃣ Home Remedies:
-
-Warm compress (garam paani se halki sinkai).
-
-Stretch your legs lightly.
-
-Agar numbness barh rahi hai toh please doctor se consult karein. Allah sehat de ❤️
-
-Use markdown formatting for clarity. Each section and point must appear on a new line using \n\n (double line break). Never merge everything into one line.
-responce shoud be like this exapmle for beterr formting and understand: **Oh, I understand you're feeling numb. Let's see what could be happening.**\n\n
- Possible Causes:\n
-- Prolonged sitting in one position *(Aik hi position mein der tak baithna)*\n
-- Poor circulation *(Khoon ki gardish mein kami)*\n
-- Nerve compression *(Asab par dabao)*\n\n
- Safe Medicine:\n
-- You can gently massage the area.\n
-- Koi bhi pain-relief balm laga sakte hain.\n\n
- Precautions:\n
-- Try to move around every 20–30 minutes. *(Har 20–30 minute baad hiley julley.)*\n
-- Maintain a good posture while sitting. *(Baithtay waqt sahih posture rakhein.)*\n\n
- Home Remedies:\n
-- Warm Compress: Garam pani ki bottle se halki sinkai karein.\n
-- Stretching: Halka warm-up karein.\n\n
- Kind Closing:\n
-Agar dard barhta hai toh please doctor ko dikhayein.\n
-Take care! ❤️
-”
-""",
-    tools=[get_info_about_health],
-    model_settings=model_settings.ModelSettings(tool_choice="required"),
-    model=model
-)
-
-async def get_health_response(user_message: str) -> str:
-    print("Running agent with message:", user_message)
-    result = await Runner.run(agent, user_message, run_config=config)
-    print("Final output:", result)
-    return result.final_output

pyproject.toml

@@ -1,22 +0,0 @@
-[project]
-name = "backend"
-version = "0.1.0"
-description = "Add your description here"
-readme = "README.md"
-requires-python = ">=3.12"
-dependencies = [
-    "eval-type-backport>=0.2.2",
-    "fastapi[standard]>=0.116.1",
-    "numpy>=2.3.1",
-    "openai-agents>=0.2.3",
-    "pillow>=11.3.0",
-    "presidio-analyzer>=2.2.359",
-    "presidio-anonymizer>=2.2.359",
-    "pypdf2>=3.0.1",
-    "python-doctr>=1.0.0",
-    "python-docx>=1.2.0",
-    "python-dotenv>=1.1.1",
-    "torch>=2.7.1",
-    "torchvision>=0.22.1",
-    "uvicorn>=0.35.0",
-]

reportanalysis.py

@@ -1,129 +0,0 @@
-
-import io
-import re
-import os
-import docx
-from PyPDF2 import PdfReader
-import numpy as np
-from PIL import Image
-from doctr.models import ocr_predictor
-from presidio_analyzer import AnalyzerEngine
-from presidio_anonymizer import AnonymizerEngine
-from dotenv import load_dotenv  
-from pydantic import BaseModel, Field
-from typing import List, Literal
-from agents import (
-        Agent,
-        AsyncOpenAI, 
-          OpenAIChatCompletionsModel, 
-        AgentOutputSchema,
-        AgentOutputSchemaBase,
-        enable_verbose_stdout_logging,
-        set_tracing_disabled
-        )
-enable_verbose_stdout_logging()
-set_tracing_disabled(True)
-
-load_dotenv()
-model = ocr_predictor(pretrained=True)
-analyzer = AnalyzerEngine()
-anonymizer = AnonymizerEngine()
-
-API = os.getenv("GEM_API_KEY")
-
-class AiInsights(BaseModel):
-    overallAssessment: str
-    keyHighlights: List[dict[str, str]]
-    dietaryRecommendations: List[str]
-    lifestyleAdvice: List[str]
-    precautions: List[str]
-    risks: List[str]
-    actions: List[dict[str, str]]
-    tips: List[str]
-
-class KeyFinding(BaseModel):
-    test: str
-    value: int
-    unit: str
-    range: str
-    shortExplaination: str
-    status: Literal["Red", "Yellow", "Green"]
-
-class AnalysisResult(BaseModel):
-  fileName: str
-  reportType: str
-  summary: str
-  keyFindings: List[KeyFinding]
-  aiInsights: AiInsights
-
-client = AsyncOpenAI(
-       api_key = API,
-       base_url = "https://generativelanguage.googleapis.com/v1beta/openai/",
-)
-
-agent_model = OpenAIChatCompletionsModel(
-        model = "gemini-2.5-flash",
-        openai_client = client,
-        
-)
-
-def format_json(result):   
-        analyzer_results = analyzer.analyze(text=result, language='en')
-        anonymized_text = anonymizer.anonymize(text=result, analyzer_results=analyzer_results)
-        result_text = anonymized_text.text
-        pattern = r'(<PERSON>\s+[\w\s\-]+)'
-        hospital_pattern = r'(?i)\b(?:[A-Z][a-zA-Z]+(?:\s+|,|&)?){1,6}(hospital|lab|clinic|diagnostic|medical|centre|pathology)\b'
-        result_text = re.sub(r'[,.()\'"-]', ' ', result_text).strip()
-        result_text = re.sub(pattern, r'<NAME>', result_text)
-        result_text = re.sub(hospital_pattern, r'<HOSPITAL>', result_text,)
-        print(result_text)
-        return result_text
-
-def extract_text(content ,pdf ,doc) -> str:
-        if pdf:
-            reader = PdfReader(io.BytesIO(content))
-            text = ''
-            for page in reader.pages:
-                text += page.extract_text() + '\n'
-            print(text)
-            return text.strip()
-        elif doc:
-            doc = docx.Document(io.BytesIO(content))
-            text = ''
-            for para in doc.paragraphs:
-                text += para.text + '\n'
-            print(text)
-            return text.strip()
-        
-        else:
-          image = Image.open(io.BytesIO(content)).convert("RGB")
-          npImg = np.ascontiguousarray(np.array(image, dtype='uint8'))
-          ORCresult = model([npImg])
-          clean_jason = format_json(ORCresult.render())
-          print(clean_jason)
-          return clean_jason
-
-
-Report_Agent = Agent(
-        name = "Report_Analysis_Agent",
-        instructions = """You are a Medical Report Analysis Agent.
-
-Your role is to analyze uploaded medical test reports and generate clear, accurate health advice in structured JSON format.
-
-Your Main Task:
-1. Analyze the extracted medical text carefully.
-2. Identify each test name, its result (user value), and the normal reference range.
-3. Assign a flag to each test based on the result:
-   - Red: Critical or abnormal
-   - Yellow: Slightly out of range or borderline
-   - Green: Normal or safe
-4. Provide a clear summary of the findings.
-5. Offer relevant AI-driven health tips, highlight potential risks, and suggest dietary and lifestyle improvements.
-6. Structure the output in the specified JSON format.
-Response format: {'type': 'json_schema', 'json_schema': {'name': 'final_output', 'strict': False, 'schema': {'$defs': {'AiInsights': {'properties': {'overallAssessment': {'title': 'Overallassessment', 'type': 'string'}, 'keyHighlights': {'items': {'additionalProperties': {'type': 'string'}, 'type': 'object'}, 'title': 'Keyhighlights', 'type': 'array'}, 'dietaryRecommendations': {'items': {'type': 'string'}, 'title': 'Dietaryrecommendations', 'type': 'array'}, 'lifestyleAdvice': {'items': {'type': 'string'}, 'title': 'Lifestyleadvice', 'type': 'array'}, 'precautions': {'items': {'type': 'string'}, 'title': 'Precautions', 'type': 'array'}, 'risks': {'items': {'type': 'string'}, 'title': 'Risks', 'type': 'array'}, 'actions': {'items': {'additionalProperties': {'type': 'string'}, 'type': 'object'}, 'title': 'Actions', 'type': 'array'}, 'tips': {'items': {'type': 'string'}, 'title': 'Tips', 'type': 'array'}}, 'required': ['overallAssessment', 'keyHighlights', 'dietaryRecommendations', 'lifestyleAdvice', 'precautions', 'risks', 'actions', 'tips'], 'title': 'AiInsights', 'type': 'object'}, 'KeyFinding': {'properties': {'test': {'title': 'Test', 'type': 'string'}, 'value': {'title': 'Value', 'type': 'integer'}, 'unit': {'title': 'Unit', 'type': 'string'}, 'range': {'title': 'Range', 'type': 'string'}, 'status': {'enum': ['Red', 'Yellow', 'Green'], 'title': 'Status', 'type': 'string'}}, 'required': ['test', 'value', 'unit', 'range', 'status'], 'title': 'KeyFinding', 'type': 'object'}}, 'properties': {'fileName': {'title': 'Filename', 'type': 'string'}, 'reportType': {'title': 'Reporttype', 'type': 'string'}, 'summary': {'title': 'Summary', 'type': 'string'}, 'keyFindings': {'items': {'$ref': '#/$defs/KeyFinding'}, 'title': 'Keyfindings', 'type': 'array'}, 'aiInsights': {'$ref': '#/$defs/AiInsights'}}, 
-'required': ['fileName', 'reportType', 'summary', 'keyFindings', 'aiInsights'], 'title': 'AnalysisResult', 'type': 'object'}}}
-""",
-        model = agent_model,
-        output_type= AgentOutputSchema(AnalysisResult, strict_json_schema=False)
-)
-

requirements.txt

@@ -1,14 +1,14 @@
-fastapi[standard]
+# For ZeroGPU, gradio/spaces/huggingface_hub/torch are preinstalled and
+# platform-managed; the Dockerfile installs them explicitly.
+diffusers>=0.33
+transformers>=4.46
+accelerate>=1.0
+gguf
+safetensors
 numpy
-openai-agents
-uvicorn
+scipy
 pillow
-presidio-analyzer
-presidio-anonymizer
-pypdf2
-python-doctr
-python-docx
-python-dotenv
-eval_type_backport
-torch
-torchvision
+sentencepiece
+protobuf
+imageio-ffmpeg
+httpx

runtime.txt

@@ -1 +0,0 @@
-python-3.10.12